Two-stage agent pipeline: the downstream agent must lock every parameter the upstream stage validated, not a subset
Stage 1 of a pipeline evaluated candidate designs defined by five parameters and selected one with evidence. Stage 2, a separate LLM step written by another person, read the selection, locked only two of the five as 'the acoustic numbers', and chose the rest itself to satisfy user needs. Running stage 2 on a real stage 1 result showed it changing two of the unlocked parameters in ways stage 1 had specifically tested and rejected, with a plausible rationale each time, so the final artifact was a design nobody evaluated. Nothing errored; both stages succeeded. Lessons: (1) the hand-off contract between stages should say which fields are validated results and must pass through unchanged, and the downstream stage should refuse or flag any change to them; (2) preferences that would alter validated parameters belong upstream as candidates to be tested, not downstream as post-hoc adjustments; (3) the only way this surfaced was diffing the upstream selected parameters against the downstream built parameters after a real run, so add that diff as an automatic check and report it. Also confirm the direction of the hand-off with the people who own each stage before building glue: I spent an hour on the reverse direction because two components could each plausibly have come first.